Systems, Methods, and Media for Detecting an Anatomical Object in a Medical Device Image

Abstract

Systems, methods, and media for detecting an anatomical object in a medical device image are provided. In some embodiments, system for detecting an anatomical object in a medical device image are provided, the systems comprising: at least one hardware processor that: applies the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed fro ma plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and identifies the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to be classifier.

Description

CROSS REFERENCE TO RELATED APPLICATION

This application claims the benefit on U.S. Provisional Patent Application No. 61/442,112, filed Feb. 11, 2011, which is hereby incorporated by reference herein in its entirety.

TECHNICAL FIELD

The disclosed subject matter relates to systems, methods, and media for detecting an anatomical object in a medical device image.

BACKGROUND

Pulmonary embolism (PE) is a relatively common cardiovascular emergency with about 600,000 cases occurring annually and causing approximately 200,000 deaths in the United States per year. A pulmonary embolus usually starts from the lower extremity, travels in the bloodstream through the heart and into the lungs, gets lodged in the pulmonary arteries, and subsequently blocks blood flow into, and oxygen exchange in, the lungs, leading to sudden death. Based on its relative location in the pulmonary arteries, an embolus may be classified into four groups (central, lobar, segmental and sub-segmental).

Computed tomography pulmonary angiography (CTPA) has become the test of choice for PE diagnosis. The interpretation of CTPA image datasets is made complex and time consuming by the intricate branching structure of the pulmonary vessels, a myriad of artifacts that may obscure or mimic PEs, and suboptimal bolus of contrast and inhomogeneity with the pulmonary arterial blood pool.

Several approaches for computer-aided diagnosis of PE in CTPA have been proposed. However, these approaches are not adequately capable of detecting central PEs, distinguishing the pulmonary artery from the vein to effectively remove any false positives from the veins, and dynamically adapting to suboptimal contrast conditions associated the CTPA scans.

Accordingly, new mechanisms for detecting an anatomical object in a medical device image are needed.

SUMMARY

Systems, methods, and media for detecting an anatomical object in a medical device image are provided. In some embodiments, system for detecting an anatomical object in. a medical device image are provided, the systems comprising: at least one hardware processor that: applies the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and identifies the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

In some embodiments, methods for detecting art anatomical object in a medical device image are provided, the methods comprising: applying the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and identifying the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

In some embodiments, non-transitory computer-readable media containing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for detecting an anatomical object in a medical device image are provided, the method comprising: applying the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and identifying the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of hardware that can be used in accordance with some embodiments.

FIG. 2 shows examples of Haar features that can be used in accordance with some embodiments.

FIG. 3 is a block diagram of a multi-stage classifier in accordance with some embodiments.

FIG. 4 is a flow diagram of a process for training a multi-stage classifier in accordance with some embodiments.

FIG. 5 is a block diagram of another multi-stage classifier in accordance with some embodiments.

FIG. 6 is a flow diagram of another process for training a multi-stage classifier in accordance with some embodiments.

FIG. 7 is a flow diagram of a process for training a single-stage classifier in accordance with some embodiments.

FIG. 8 is a flow diagram of a process for detecting objects in images using a classifier in accordance with some embodiments.

DETAILED DESCRIPTION

Systems, methods, and media for detecting an anatomical object in a medical device image are provided. More particularly, in some embodiments, systems, methods, and media for detecting an anatomical object, such as a pulmonary trunk, in a medical device image, such as a computed tomography pulmonary angiography (CTPA) image, are provided.

The pulmonary trunk is the main pulmonary artery that rises from the right ventricle of the heart, extends upward, and divides into the right and left pulmonary arteries carrying blood to the lungs. Because PEs are only found in the pulmonary artery, identifying the pulmonary trunk in medical device images, such as CTPA images, can be used in PE diagnosis.

Turning to FIG. 1, an example of hardware 100 that can be used in accordance with some embodiments is illustrated. As shown, this hardware can include an imaging device 102 and an image processing device 104. Imaging device 102 can be any suitable device for generating imaging data that can be provided to image processing device 104. For example, in some embodiments, imaging device 102 can be a computed tomography (CT) scanner. Image processing device 104 can be any suitable device for receiving and processing imaging data. For example, in some embodiments, image processing device 104 can be a computer. Imaging device 102 can communicate with image processing device 104 in any suitable manner such as via a direct connection between the devices, via a communication network, etc.

In some embodiments, image processing device 104 can be any of a general purpose device such as a computer or a special purpose device such as a client, a server, etc. Any of these general or special purpose devices can include any suitable components such as a hardware processor (which can be a microprocessor, digital signal processor, a controller, etc), memory, communication interfaces, display controllers, input devices, etc.

In some embodiments, imaging device 102 and image processing device 104 can be integrated into a single device.

In some embodiments, a machine-learning-based approach can be used by image processing device 104 for automatically detecting an anatomical object, such as a pulmonary trunk, in a medical device image.

More particularly, for example, in some embodiments, a cascaded AdaBoost classifier can be trained with a large number of Haar features (example of which are shown in FIG. 2) extracted from computed tomography pulmonary angiography (CTPA) image samples, so that an anatomical object, such as a pulmonary trunk, can subsequently be automatically identified by sequentially scanning CTPA images and classifying each encountered sub-image with the trained classifier. In some embodiments, CTPA images can be automatically scanned at multiple scales to handle size variations of the anatomical objects (e.g., pulmonary trunks).

An AdaBoost classifier is a type of machine learning algorithm drat combines weak learners to create a single strong learner. A weak learner is a classifier that may perform only slightly better than random guessing. A commonly used weak classifier called the decision stump can be used to make a prediction based on the value of a single input feature.

For example, h₁, h₂, . . . , h_N make up a set of weak learners, a combination of these weak learners can be written as:

F(x)=Σ_j−1^Nf_j(x)=Σ_j=1^Nω_jh_j(x),

where ω_j is the corresponding coefficient for weak learner h_j. Boosting is a process to select weak learners h_j and determine their coefficients ω_j, so as to combine the selected weak learners to form a strong learner F(x).

In some embodiments, AdaBoost can he used to select the most relevant, features from any suitable number (e.g., thousands) of Haar features, each corresponding to a weak learner. In some embodiments, a Haar feature can be defined in terms of two adjacent rectangle regions, which can be illustrated in white and black as shown in FIG. 2, for example. The value of a Haar feature can be the sum of any suitable pixels values (such as intensity) in one or more first rectangle(s) (e.g., the white rectangles) of the feature minus the sum of the suitable pixel values in one or more second rectangle(s) (e.g., the black rectangle(s)) of the feature.

In some embodiments, any suitable criteria, such as desired true positive rate, false positive rate, and number of weak learners, can be used to determine the number of strong boosted classifiers, the number of weak learners in each boosted classifier, and the relative operating characteristic (ROC) operating points (which can can be selected from a ROC curve produced during training) for classifying images. For example, in some embodiments, a True Positive Rate (TPR) a, a False Positive Rate (FPR) β_i, and a maximum number of weak learners η_i can be used as criteria for training a cascaded classifier stage.

As shown in FIG, 3, an AdaBoost classifier 300 can include any suitable number of strong classifier stages 302, 304, and 306. D_i⁺, D_i⁻ can be used to refer to positive sub-images and negative sub-images that can be used for training an AdaBoost classifier stage i. In each stage 302, 304, or 306, during training, weak learners can be added to tire stage until a given target performance (α_i, β_i) or a given number of weak learners η_i in the stage is reached. The output of the training at stage i is a boosted classifier containing weak learners from f_τi−1+1 to f_96i. Upon completing training a given stage, new negative samples can be classified by the stage to identify false positives (i.e., negative samples which are classified as positive) and then these negative samples (which are falsely classified as positives) can be combined with the negative samples used for training the current stage and the combination used for training the subsequent stage.

Turning to FIG. 4, an example process 400 for training this classifier in

accordance with some embodiments is shown. As illustrated, after process 400 begins at 402, the process selects a first stage of the classifier to train. This stage can be selected in any suitable manner. Next, at 406, the process can select an initial set of weak learners for the stage. Any suitable number of weak learners, including one, can be selected, and the weak learners can be selected in any suitable manner, such as randomly. Then, at 408, process 400 can apply positive and negative sub-image samples to the set of weak learners. Any suitable number of positive and negative sub-image samples (e.g., 100 each) can be applied, and these samples can be selected for application in any suitable manner, such as randomly. The process can then determine at 410 whether the performance of the stage is sufficient or whether the maximum number of weak learners for the stage has been reached. Any suitable criteria or criterion can be used for determining whether the performance of the stage is sufficient in some embodiments. For example, in some embodiments, the performance of the stage can be deemed to be sufficient when the TPR α_i is over 0.99 and FPR β_i is below 0.05. Any suitable threshold η_i for a maximum number of weak learners can be used in some embodiments. For example, η_i can be 30 in some embodiments. If it is determined at 410 that the performance is not sufficient and the maximum number of weak learners has not been reached, then process 400 can add one or more weak learners to the set at 412 and loop back to 408. The weak learners to be added can be selected in any suitable manner (e.g., randomly) and any suitable number of weak learners (including one) can be added, in some embodiments. Otherwise, at 414 process 400 can then assign the set of weak, learners to the boosted strong classifier for the current stage. Next, at 416, process 400 can use the set of weak, learners to detect new negative samples that appear positive (i.e., false positives) and add these new negative samples to the set of negative samples and use this new set for the next stage. Any suitable number of new negative samples, such as 100. can be used in some embodiments. At 418, process 400 can then determine whether the current stage is the last stage, and, if not, select the next stage at 420. Otherwise, process can end at 422.

Another example classifier 500 that can be used in some embodiments is illustrated in FIG. 5. As shown, classifier 500 can include any suitable number of strong classifier stages 502, 504, and 506. D_i⁺, D_i⁻ can be used to refer to positive sub-images and negative sub-images that can be used for the teaming a classifier stage i. In each stage 502, 504, or 506, during training, weak learners can be added to the stage until a given target performance (α_i, β_i) or a given number of weak learners η_i in the stage is reached. The output of the training at stage i is a boosted classifier 504 containing weak learners from f₁ to f_96i. That is, a stage can include all of the weak learners of all previous stages in some embodiments. Upon completing training a given stage, new negative samples can be classified by the stage to identify false positives (i.e., negative samples which are classified as positive) and then these negative samples (which are falsely classified as positives) can be added to the negative samples from the current stage and used for training the subsequent stage.

Turning to FIG. 6, an example process 600 for training this classifier in accordance with some embodiments is shown. As illustrated, process 600 includes steps 402, 404, 406, 408,410, 412, 414, 416, 418, 420, and 422 which can be performed as described above in connection with process 400 of FIG. 4. Unlike in process 400, however, after performing step 420, process 600 can branch to step 412 rather than step 406.

In some embodiments, rather than using a multi-stage classifier as described above, a single stage classifier can be used. Such a classifier may include a single classifier stage 302 as shown in FIG. 3.

Turning to FIG. 7, an example process 700 for training this classifier in accordance with some embodiments is shown. As illustrated, process 700 includes steps 402, 406, 408, 410, 412, 414, and 422 which can be performed as described above in connection with process 400 of FIG. 4. However, unlike process 400, in this approach, 100 positive samples (or any other suitable number) and 500 negative samples (or any other suitable number) can be used to train the single stage, and training can be completed when the TPR α_i=100, the FPR β_i=0, and when the number of weak classifiers η_i=100. In some embodiments, negative samples can be false positive samples from other training techniques as described above. (00351 As described above, to perform detection using a classifier, an image can be

provided to the one or more stages of the classifier and a positive indication or a negative indication can be provided. If at any stage in the classifier, an image is classified as negative, the image can be removed from subsequent testing by subsequent stages of the classifier and the classification of the image can be maintained as negative.

Turning to FIG. 8, an example process 800 for detecting images in accordance with some embodiments is shown. As illustrated, after process 800 begins at 802, the process can select a detection scheme at 806. Any suitable detection scheme can be used, such as the multi-stage or single-stage schemes described above. Next, the first image can be selected at 806. The first image can be selected in any suitable manner (e.g., such as randomly, in-order, etc), and the image can be any suitable portion of another image (e.g., such as a random portion of a first image). Then, at 808, the first strong classifier in the selected scheme can be selected. At 810, the image can then be applied to the selected strong classifier, which can assign a classification and a score to the image. At 812, process 800 can then determine if the classification from the stage is negative. If so, the next image can be selected at 814 and process 800 can loop back to 808. Otherwise, at 816, it can be determined if the current stage is the last strong classifier. If not, then process 800 can select the next strong classifier at 818 and loop back, to 810. Otherwise, process 800 can classify the image as positive at 820 and merge the image with any previous overlapping, positive-classified images at 822. Any suitable images can be. identified as being overlapping in some embodiments. For example, images can be identified as being overlapping if the images share over 25% of their data (e.g., based, on location and size of the image) and/or if their z-axis distance is less than five pixels, in some embodiments, when merging images, their individual scores can be added together. Next, at 824, process 800 can determine if the current image is the last image. If not, the process can select the next image at 814 and loop back to 808. Otherwise, the process can select the highest-score merged image as the detected object at 826 and terminate at 828.

In some embodiments, any suitable computer readable media can be used for storing instructions for performing the processes described herein, such as performing training of classifiers and classifying of images. For example, in some embodiments, computer readable media can be transitory or non-transitory. For example, non-transitory computer readable media can include media such as magnetic media (such as hard disks, floppy disks, etc.), optical media (such as compact discs, digital video discs, Blu-ray discs, etc.), semiconductor media (such as flash memory, electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), etc.), any suitable media that is not fleeting or devoid of any semblance of permanence during transmission, and/or any suitable tangible media. As another example, transitory computer readable media can include signals on networks, in wires, conductors, optical fibers, circuits, any suitable media that is fleeting and devoid of any semblance of permanence during transmission, and/or any suitable intangible media.

Although the invention has been described and illustrated in the foregoing illustrative embodiments, it is understood that the present disclosure has been made only by way of example, and that numerous changes in the details of implementation of the invention can be made without departing from the spirit and scope of the invention, which is limited only by the claims that follow. Features of the disclosed embodiments cm be combined and rearranged in various ways.

Claims

1. A system for detecting an anatomical object in a medical device image, comprising:

at least one hardware processor that: applies the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and identifies the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

2. The system of claim 1, wherein the classifier is an AdaBoost classifier.

3. The system of claim 1, wherein the classifier is based on a training of the classifier using positive samples and negative samples for the anatomical object.

4. The system of claim 1, wherein the anatomical object is a pulmonary trunk.

5. The system of claim J, wherein the at least one hardware processor also merges the medical device image with at least one other medical device image.

6. The system of claim 1, wherein the at least one hardware processor also determines that the medical device image has the highest response of a plurality of medical device images.

7. The system of claim 1, wherein the medical device image is a computed tomography pulmonary angiography image.

8. A method for detecting an anatomical object in a medical device image, comprising:

applying the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and

identifying the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

9. The method of claim 8, wherein the classifier is an AdaBoost classifier.

10. The method of claim 8, wherein the classifier is based on a training of the classifier using positive samples and negative samples for the anatomical object.

11. The method of claim 8, wherein the anatomical object is a pulmonary trunk.

12. The method of claim 8, further comprising merging the medical device image with at least one other medical device image.

13. The method of claim 8, further comprising determining that the medical device image has the highest response of a plurality of medical device images.

14. The method of claim 8, wherein the medical device image is a computed tomography pulmonary angiography image.

15. A non-transitory computer-readable medium containing computer-executable instructions that, when executed by a processor, cause the processor to perform a method for detecting an anatomical object in a medical device image, the method comprising:

applying the medical device image to a classifier having a plurality of stages, wherein a first stage of the plurality of stages and a second stage of the plurality of stages each includes a strong learner formed, from a plurality of weak learners, and the weak learners in the second stage include a plurality of the weak learners included in the first stage; and

identifying the medical device image as being positive or negative of showing the anatomical object based on the application the medical device image to the classifier.

16. The non-transitory computer-readable medium of claim 15, wherein the classifier is an AdaBoost classifier.

17. The non-transitory computer-readable medium of claim 15, wherein the classifier is based on a training of the classifier using positive samples and negative samples for the anatomical object.

18. The non-transitory computer-readable medium of claim 15, wherein the anatomical object is a pulmonary trunk.

19. The non-transitory computer-readable medium of claim 15, wherein the method further comprises merging the medical device image with at least one other medical device image.

20. The non-transitory computer-readable medium of claim 15, wherein the method further comprises determining that the medical device image has the highest response of a plurality of medical device images.

21. The non-transitory computer-readable medium of claim 15, wherein the medical device image is a computed tomography pulmonary angiography image.